The Agent Skills Directory

[SAFE]: The skill implements a structured workflow for manuscript editing with built-in verification steps to ensure data integrity. The local Python scripts for fidelity and sentence variety checks are used for quality assurance and do not exhibit malicious behavior or unauthorized file access.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided manuscripts, which constitutes an attack surface for indirect prompt injection. However, the risk is mitigated by explicit instructions and programmatic checks. 1. Ingestion points: Manuscript text files or sections read from the user environment (SKILL.md, Phase 1). 2. Boundary markers: While the skill lacks explicit delimiters for untrusted content, the agent is instructed to focus on scanning for specific linguistic patterns rather than executing the text. 3. Capability inventory: The skill uses Read, Write, Edit, Grep, Glob, and Bash tools. The Bash tool is used to execute local verification scripts. 4. Sanitization: The check_rewrite_fidelity.py script acts as a safety check by verifying that the agent's edits have not altered numeric values or dropped citations, providing a technical invariant that prevents malicious instructions in the manuscript from causing silent data corruption.

humanize